Errors and limits

Each dialect reports failures in its own error shape. The OpenAI routes return an error object; the Anthropic routes return a type: "error" envelope. The status code means the same thing in both.

Error bodies

{
  "error": {
    "message": "The model does not exist or you do not have access to it.",
    "type": "invalid_request_error",
    "param": null,
    "code": "model_not_found"
  }
}

{
  "type": "error",
  "error": {
    "type": "not_found_error",
    "message": "The model does not exist or you do not have access to it."
  }
}

OpenAI bodies carry code; Anthropic bodies carry only type and message. Match on the status and message, not on an exact type string, when you need code that works across both.

Statuses

Status	OpenAI `code`	Anthropic `type`	Message	Cause
`400`	`invalid_model`	`invalid_request_error`	`"model" must be a non-empty string.`	The body has no usable `model`.
`400`	`invalid_stream`	`invalid_request_error`	`"stream" must be a boolean.`	`stream` was not a boolean.
`400`	`invalid_messages`	`invalid_request_error`	`"messages" must be a non-empty array.`	Chat or Messages request without messages.
`400`	`invalid_prompt`	`invalid_request_error`	`"prompt" is required.`	Legacy completions request without a prompt.
`400`	`invalid_input`	`invalid_request_error`	`"input" is required.`	Responses request without an input.
`400`	`stream_not_supported`	`invalid_request_error`	`This model does not support streaming responses.`	`stream: true` on a model configured without streaming.
`401`	`invalid_api_key`	`authentication_error`	`Missing API key.` or `Incorrect API key provided.`	No key, or a key that is unknown or deleted.
`404`	`model_not_found`	`not_found_error`	`The model does not exist or you do not have access to it.`	Wrong id, paused or draft model, or access that ended.
`429`	`rate_limit_exceeded`	`rate_limit_error`	`Rate limit exceeded for this model.`	The model’s rate limit, protection at Noket, or a budget the gateway refuses.
`429`	`insufficient_quota`	`rate_limit_error`	`Your traffic quota has been exceeded.`	The model’s traffic (byte) quota is used up.
`500`	`internal_error`	`api_error`	`Internal server error.`	Noket failed before or after the provider call.
`502`	`upstream_error`	`api_error`	`The upstream provider failed to complete the request.`	The provider returned an error Noket cannot attribute more precisely.
`502`	`upstream_auth_error`	`api_error`	`The upstream provider rejected the model credentials.`	The provider refused the stored key or account.
`504`	`upstream_timeout`	`timeout_error`	`The upstream provider did not respond in time.`	The provider did not answer within the model’s timeout.

A 429 from an exhausted budget and one from a rate limit share the message Rate limit exceeded for this model. The console distinguishes them: Usage shows spend against each budget, and the model page shows its current state.

How upstream failures are mapped

When the provider behind a model fails, Noket translates its status instead of forwarding its body. Provider messages, credentials, and internal model names never reach the client.

Provider response	Noket returns
`400`	`400` with `upstream_invalid_request`
`401` or `403`	`502` with `upstream_auth_error`
`404`	`404` with `model_not_found`
`429`	`429` with `rate_limit_exceeded`
Timeout	`504` with `upstream_timeout`
Anything else	`502` with `upstream_error`

An upstream_auth_error almost always means the model’s credentials need attention in the console or, for a tunnel model, that the connector holding the key is offline or misconfigured. Ask the model’s owner; if that is you, check the upstream in noket upstream test.

Errors during a stream

Once a response has started streaming, the status is already 200 and cannot change. If the provider fails mid-stream, the stream ends, and the provider’s error event arrives in the stream. Treat an incomplete stream — missing [DONE], or an error event — as a failed call. The request appears in the model’s Logs, though a failed or cancelled stream may have no reported cost.

Retries

Safe retry statuses are 429, 502, 503, 504, and connection failures. Use exponential backoff with jitter. Do not retry 400, 401, or 404 unchanged — the same request will fail again. A streamed request is not safe to retry after partial output unless the client can drop the partial response.